DLBacktrace
Deep models span many architectures and tasks, which makes relevance tracing difficult to inspect consistently. DLBacktrace is a model-agnostic explainability framework for transformers, language models, and Mixture-of-Experts architectures.
View repositoryArchitecture & implementation
I contributed to graph capture, activation collection, relevance propagation, trace storage, and visualization. For LLaMA-family and MoE support, I used PyTorch torch.export/FX graphs and implemented routing-aware relevance propagation. The repository presents this as a model-agnostic framework across transformers, language models, MoE, and other task types.
Design constraint — tracing must cover model execution and relevance flow across different architectures; a single-model explanation path would not meet the framework's stated scope.
Inspect the implementation
Status & limits — the public repository contains the framework implementation. The résumé lists comparisons with SHAP, LIME, and Integrated Gradients, but no benchmark figures are repeated here; public README performance claims are not independently reproduced on this site.